bytedance/deer-flow

What it solves

DeerFlow is a "super agent harness" designed to orchestrate complex, multi-step research and task execution. It solves the problem of managing multiple specialized sub-agents, long-term memory, and secure execution environments (sandboxes) to perform deep exploration and efficient research flows.

How it works

DeerFlow acts as an orchestration layer that combines several core components:

• Sub-Agents: It can spawn and manage multiple agents to handle different parts of a task.
• Skills & Tools: It uses extensible skills, including integrations with tools like Claude Code and InfoQuest for intelligent search and crawling.
• Sandboxes: It provides a secure file system and execution environment to run code and manage files without risking the host system.
• Memory & Context: It employs long-term memory and "context engineering" (including manual context compaction) to maintain coherence over long research sessions.
• Model Flexibility: It supports a wide range of LLM providers (OpenAI, Anthropic, vLLM, OpenRouter) and specialized reasoning models.

Who it’s for

It is intended for developers and researchers who need an autonomous agent system capable of performing deep research, coding tasks, and complex explorations that require tool use and execution in a controlled environment.

Highlights

• Super Agent Architecture: Orchestrates sub-agents, memory, and sandboxes.
• Integrated Search: Built-in support for BytePlus InfoQuest for advanced web and image search.
• ** uma Sandbox Mode**: Secure execution of bash commands and file writes.
• Extensible Tooling: Integration with MCP servers and various LLM providers.
• Comprehensive Tracing: Supports LangSmith, Langfuse, and Monocle for debugging and performance monitoring.